Processing GHG emissions disclosures

I pursue a 2 x 2 between-subjects experimental design with the following two interventions:
* First, I randomize whether participants see sustainability reports published before (i.e., for the fiscal year 2022) or after (fiscal year 2024) the CSRD took effect. As reports published for the fiscal year did not include any requirement from the Directive, these constitute a plausible baseline. I ensure, however, that reports from both years contain the same information. This means that they should only differ in the way the information is presented due to standardization.
* Second, I randomize whether or not participants have access to an AI tool designed to extract information from the reports given the increasing role of AI in analyzing large, unstructured datasets. The AI tool is built specifically for the purpose of retrieving information from corporate reports and leverages a Retrieval Augmented Generation (RAG) architecture.

2025-07-01

2025-09-30

To conceptualize processing costs, I rely on the framework proposed by Blankespoor (2019). Specifically, I focus on the latter two stages, acquisition and integration costs:
* My measure of acquisition costs is the time it takes participant p to extract the relevant information from all reports in industry r.
* My measure of integration costs is based on participant p’s understanding of industry r’s emissions performance and is computed as the sum of questions answered correctly for industry r.

* Ranking dispersion
* Learning
* Querying behavior
* Task satisfaction
* Report engagement

Participants get access to four firms' annual and/or sustainability reports over a custom website. On the website, they can collect and compare quantitative indicators and take notes. They then have to answer questions on these reports.

After entering my custom experimental platform, participants read the task description and see an overview that lists four firms from one of two randomly selected industries and the five single-choice questions on which they will be evaluated. Questions relate to absolute and relative emissions, emissions intensity, emissions reduction technologies and emissions targets. Before they start the analysis, I elicit participants’ prior perception of the firms’ ESG performance by letting them rank the companies.

The overview contains a link to a custom subpage that displays each firm’s sustainability report and a table for entering quantitative indicators, along with space for notes. As long as they have not started the assessment task (which starts automatically after 30 minutes, but can be started earlier), participants can access the reports and the AI tool. Once they start the assessment task, participants will have to answer the five questions only relying on their extracted indicators and notes. Having finished the assessment for one industry, participants are asked to rank the companies again allowing me to elicit any changes from their prior assessment, before they are redirected to the second and final industry. To obtain their certificate of completion, they have to answer ten demographical and background questions.

Assignment to a group inside a treatment arm is independently done on the server when participants log in to the platform.

across-participant randomization of two treatment arms (i.e., four groups in total)

No

-

>120 participants

>30

-